Bones need to be placed at the centers of rotation, rather than the exact bone locations.
The place where this is the biggest issue is with the spine. Your organs can only compress/extend so much, so when you bend forward, your vertebrae kinda expand or something (they’re separated by more flexibles discs). What happens is that you don’t literally rotate around the position of the spine bones, but much more centrally than that.
You can see this pretty easily by trying to rig a model and placing the spine realistically. Arching the back will make the front of your torso seem to expand unnaturally. In reality, there are actually a lot of different things going on, it’s not a simple rotation about some point, but the closest simple abstraction is that your body is rotating about some point somewhat forward (“anterior”) of your spine.
Other places, the center of rotation is at the location of the bone. Compression isn’t a big issue for knees, and the ankle ends up maintaining a constant distance from roughly the center of the leg. (Not necessarily true near the joint itself though.)
The actual directions that bones point doesn’t really matter-- Blender’s “tails” don’t matter. They’re used to establish autoweights and they’re used for local axes, and they can be used by some constraints. But the only important thing for the final model is where the head is. One of the things that means is that you should feel to rotate tails however you want to get the autoweights you want. So should your spine rotate with your model? If your model is modelled with a straight back, your spine bones should be in a straight line! If your model is not (modelled with near-porn back arch, say) then your spine bones are not going to be in a line. The important thing is where bones are rotating about, relative to how the mesh has been modelled.
Now, if you look at an actual human being, and ask them to wiggle their jaw, what you’ll find is that the jaw maintains a roughly constant distance from the temporomandibular articulation. So for a realistic model, that’s where the jaw needs to rotate about, or that’s where a parent of distal child bones like “chin” needs to rotate about. That doesn’t mean that autoweights from a bone running from the temporomandibular to the chin will give you good autoweights near the joint itself. That’s a different issue.
But the way I think about it is, you don’t see bones on character models. You see flesh. This whole idea of “bone” is just a metaphor. That’s why you don’t model forearms as an ulna/radius pair. You don’t care about the ulna and the radius. You care about the muscles which are attached to the bones, the skin that is stretched over those muscles, and that has different behavior than the bones themselves. So you don’t have twin long bones, you have short twist bones.
For bone placement relative to the surface of the skin, the important thing to recognize is first, that armatures are merely simplified abstractions; second, that bone placement for autoweights and for posing are different things. In reality, the clavicle connects the scapula and the sternum. It does not control all shoulder movement; the scapula is much more influential. But how many models have scapulae? Few. The entire model is controlled by a simplified scapula/rib/humerus/clavicle structure. Where should that rotate about? In reality, you’d have multiple rotation points and complicated constraints connecting them. In reality? The scapula is generally ignored, while more care is taken for the clavicle, which rotates about the articulation with the sternum, more anterior than posterior. Mostly, not because that’s real, but because we don’t care about our model’s back-- we’re not looking at it very often. (And in my experience, getting good scapula deformation is a real chore. It’s one of those places where there’s a lot of skin sliding, which skeletal deformation isn’t built to simulate.)
But autoweights are something different. How do they work? The closer the line representing the bone is to the surface of the mesh, the sharper the weights are going to be. They’re going to work best if you maintain the origin of the bone, but you can do autoweights and then shift the tail afterwards to your heart’s content, so you can manipulate the sharpness of weights. This makes some sense: if you’re placing your clavicle, place it near the skin and it will mimic a superficial clavicle deforming the skin, while it will deform the scapula only very generally and loosely, in conjunction with spine and arm bones.
All that is just my experience. I don’t want to pretend like I’m an expert rigger. I have plenty to improve. Don’t treat it like it’s authoritative.
